1 Introduction

This tutorial summarizes details of the Bayesian Joint Spike-and-slab Graphical Lasso (SSJGL) algorithm (Li et. al 2019), then walks through the steps needed to run the spikeyglass package using a real-world metabolomics dataset.

1.1 Important Notes

Code throughout the document is “hidden” but if you click on “CODE” in the upper right above any R output, you can expand it to view code and (extensive) comments!

All code is available on my GitHub.

To run this tutorial on your computer please either clone the entire repository to your computer, double click ssjgl-tutorial.Rproj, open the .Rmd file and run the code as is after making sure you have installed the appropriate packages in R (below).

You can alternatively download the .csv and .Rmd files here and run the script after installing appropriate packages (below, first code chunk).

knitr::opts_chunk$set(echo = TRUE, message = FALSE, warning = FALSE)
library(tidyverse)
#devtools::install_github("mljaniczek/spikeyglass")
library(spikeyglass)
library(igraph)
library(RColorBrewer)

source("helpers.R") #custom helper functions for pretty igraph plots
theme_set(theme_bw()) #setting ggplot theme 
igraph_options(annotate.plot = TRUE) # setting default to annotate plot with number of vertices and edges
# read in metabolomics data from Raji Balasubramanian, which is on the github for this tutorial
#http://github.com/mljaniczek/ssjgl-tutorial
dat <- read.csv("hapo_metabolomics_2020.csv")

1.2 Data description

Data comes from Hyperglycemia and Adverse Pregnancy Outcome (HAPO) Study. Metabolites were measured from blood samples taken from pregnant women from four ancestry groups: Afro-Caribbean, Mexican American, Northern European, and Thai.

The data consists of ID variable, ancestry group, and metabolites (p = 51 features). The ancestry groups (K = 4) are balanced (n = 400 each, with 1600 total observations).There is also a variable “fpg” which is fasting plasma glucose, which I omit from this analysis.

Within the data there are three groups of metabolites: Amino Acids, Acycl carnitines, and Other.

Here’s what the raw data look like:

# examine ancestry group variable
#summary(as.factor(dat$anc_gp))

head(dat)[,1:10]
##       id anc_gp  fpg    mt1_1     mt1_2    mt1_3    mt1_4     mt1_5    mt1_6
## 1 hm0001    ag3 75.6 218.2223  76.99525 19.06366 14.23091  86.75162 135.2109
## 2 hm0002    ag3 84.6 292.6314 136.41320 43.14854 17.77549 120.17344 213.6531
## 3 hm0003    ag4 79.2 361.1135  79.98370 22.15848 13.05497  74.75441 136.1587
## 4 hm0004    ag3 73.8 274.0428  79.76154 19.72682 12.68744  69.34854 146.7262
## 5 hm0005    ag1 91.8 270.9049  88.98260 39.20949 15.73498  95.56857 145.2876
## 6 hm0006    ag4 73.8 271.4034  99.63992 27.81137 20.91344  95.55926 223.7438
##      mt1_7
## 1 64.00578
## 2 91.30156
## 3 83.67878
## 4 72.67280
## 5 48.49431
## 6 68.94172

We can take a look at a few of the metabolites to get an idea of distribution and missingness.

summary(dat$mt1_1)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##   115.6   281.8   322.4   328.5   370.6   660.4
summary(dat$mt2_1)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
##  -4.631  -3.341  -3.040  -3.044  -2.750  -1.154       1
summary(dat$mt3_1)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max.    NA's 
##   17.28   21.39   22.64   22.96   24.64   27.25      11

1.3 Data Processing

1.3.1 Center and Scale Data within Class

Assumptions of SSJGL need data to be centered and scaled so we must center and scale each metabolite feature.

Here I am making the data into “long” format and standardizing each metabolite separately. Check out how the data look now:

# prepare data into list of K datasets as needed as input to ssjgl function

# make data into long format then center and scale by metabolite

## we do not want to center and scale within each group because that would remove any signal we see, right? 

dat_long <- dat %>%
  select(-fpg) %>%
  pivot_longer(cols = starts_with("mt"),
               names_to = "metabolite",
               values_to = "value",
               values_drop_na = TRUE) %>%
  group_by(metabolite) %>%
  mutate(scaled_value = scale(value, center = TRUE, scale = TRUE)) %>%
  ungroup()

head(dat_long)
## # A tibble: 6 × 5
##   id     anc_gp metabolite value scaled_value[,1]
##   <chr>  <chr>  <chr>      <dbl>            <dbl>
## 1 hm0001 ag3    mt1_1      218.            -1.66 
## 2 hm0001 ag3    mt1_2       77.0           -0.942
## 3 hm0001 ag3    mt1_3       19.1           -1.30 
## 4 hm0001 ag3    mt1_4       14.2           -0.744
## 5 hm0001 ag3    mt1_5       86.8           -0.591
## 6 hm0001 ag3    mt1_6      135.            -1.17

Now we can check on how the distribution of a few of the variables look now. Data looks roughly normal. Values of metabolites in class 3 (“Other”) appear to have variable mean between ancestry groups.

# if you want to look at one metabolite
# dat_long %>% 
#   filter(metabolite %in% c("mt1_1")) %>%
#   ggplot(aes(x=value, color=anc_gp)) +
#     geom_density()+
#   facet_grid(metabolite ~.)

# demonstrative density plot showing centering/scaling of data
dat_long %>% 
  filter(metabolite %in% c("mt1_1", "mt2_1", "mt3_1")) %>%
  ggplot(aes(x=scaled_value, color=anc_gp)) +
    geom_density()+
  facet_grid(metabolite ~.)

1.3.2 Make K n by p matrices and deal with missingness

Now the raw data is centered and scaled! But first we need to get it into K n by p matrices as required by the spikeyglass package.

So in our case, we need 4 matrices (4 ancestry groups), each with 400 observations of 51 metabolites.

Note: Normally we would impute any missing data. For simplicity in this code I treated missing data as “0” value.

# make list of ancestry groups
ancestry_groups <- sort(unique(dat$anc_gp))

# filter by ancestry group, then make wide matrix of scaled values
# create a list of the results
# Here I'm using the `purrr::map` to iterate over the 4 ancestry groups and make the list. You could use base R lapply instead if you prefer.
dat_mat_list <- purrr::map(ancestry_groups,
                    ~dat_long %>%
                      filter(anc_gp == .x) %>%
                      select(-c(anc_gp, value)) %>%
                      pivot_wider(names_from = metabolite, 
                                  values_from = scaled_value,
                                  #for simplicity putting 0 as missing values. Normally should do imputation.
                                  values_fill = 0) %>%
                      select(-id) %>%
                      as.matrix())

names(dat_mat_list) <- ancestry_groups

Let’s check to see if our list contains the right structure:

str(dat_mat_list)
## List of 4
##  $ ag1: num [1:400, 1:51] -0.869 -0.474 -0.893 1.082 0.578 ...
##   ..- attr(*, "dimnames")=List of 2
##   .. ..$ : NULL
##   .. ..$ : chr [1:51] "mt1_1" "mt1_2" "mt1_3" "mt1_4" ...
##  $ ag2: num [1:400, 1:51] 0.169 0.698 -1.174 -1.012 -0.591 ...
##   ..- attr(*, "dimnames")=List of 2
##   .. ..$ : NULL
##   .. ..$ : chr [1:51] "mt1_1" "mt1_2" "mt1_3" "mt1_4" ...
##  $ ag3: num [1:400, 1:51] -1.664 -0.541 -0.822 0.152 -0.591 ...
##   ..- attr(*, "dimnames")=List of 2
##   .. ..$ : NULL
##   .. ..$ : chr [1:51] "mt1_1" "mt1_2" "mt1_3" "mt1_4" ...
##  $ ag4: num [1:400, 1:51] 0.4919 -0.8616 0.5032 0.0342 -0.157 ...
##   ..- attr(*, "dimnames")=List of 2
##   .. ..$ : NULL
##   .. ..$ : chr [1:51] "mt1_1" "mt1_2" "mt1_3" "mt1_4" ...

It does! So now we have a list of 4 matrices, each with 400 standardized observations of 51 metabolites.

1.3.3 Heatmap to visualize the processed data

Before we proceed, let’s visualize the processed data quick in a heatmap to get an idea of if the data has any patterns.

# making one full wide data frame with all 1600 observations, using the centered-scaled values for metabolites that we processed above. 
met_wide_all <- dat_long %>%
  select(-value) %>%
  pivot_wider(names_from = metabolite, 
              values_from = scaled_value,
              #for simplicity putting 0 as missing values. Normally should do imputation.
              values_fill = 0) %>%
  arrange(anc_gp) #arranging by ancestry group for the heatmap

#make matrix out of dataframe, removing the id and group labels since pheatmap just takes numeric matrix as input
met_wide_all_mat <- t(as.matrix(met_wide_all%>%
  select(-c(id, anc_gp))))
met_wide_all_mat[is.na(met_wide_all_mat)] <- 0
colnames(met_wide_all_mat) <- met_wide_all$id

#pheatmap can annotate your heatmap nicely with colors per class! so I'm preparing those here. 

#first I want the columns annotated by ancestry group
my_sample_col <- data.frame(ancestry = as.factor(met_wide_all$anc_gp))
rownames(my_sample_col) <- met_wide_all$id
# then I want the rows annotated by metabolite group, we have three groups of metabolites mt1, mt2, mt3
met_group <- str_split_fixed(rownames(met_wide_all_mat), "_", 2)[,1] %>%
  str_replace_all(c("mt1" = "Amino Acids",
                  "mt2" = "Acyl carnitines",
                  "mt3" = "Other"))
my_sample_row <- data.frame(met_group = met_group)
row.names(my_sample_row) <- row.names(met_wide_all_mat)

# finally we input the numeric matrix and annotations
pheatmap::pheatmap(met_wide_all_mat, 
         cluster_cols = FALSE, 
         show_colnames = FALSE, 
         show_rownames = FALSE, 
         annotation_col = my_sample_col,
         annotation_row = my_sample_row)

Here the x axis has the 1600 samples, sorted by ancestry group. The y axis has the 51 metabolites, clustered using hierarchical clustering, and with a color label based on metabolite group (there are 3 “groups” in the data, with suffix mt1, mt2, and mt3 which indicate Amino Acids, Acyl carnitines, and other groups respectively).

From this heatmap we can see there are big differences in the values for “Other” metabolite group (pink) across the 4 ancestry groups. metabolites from Amino Acids and Acyl Carnitines groups appear to be more homogeneous across the classes.

2 Application of SSJGL to metabolomics dataset

Let’s get to bayesian joint spike-and-slab graphical lasso now! We will essentially be taking the data from each ancestry section simultaneously perform variable selection and shrinkage by using the doubly-spike-and-slab extensions to Danaher’s 2014 GGL and FGL.

2.1 Load spikeyglass

I created the spikeyglass package using scripts provided with the paper:

Li ZR, McCormick TH, Clark SJ. Bayesian Joint Spike-and-Slab Graphical Lasso. Proc Mach Learn Res. 2019 Jun;97:3877-3885. PMID: 33521648; PMCID: PMC7845917.

The package allows us to mirror the same analysis as the frequentist Fused Graphical Lasso method in the JGL package.

library(spikeyglass)

2.2 Running DSS-FGL

We will first use the spikeyglass package to run doubly spike-and-slab fused graphical lasso (DSS-FGL) to jointly estimate the graphs in our 4 ancestry groups.

2.2.1 Set parameters

Per the SSJGL paper, the converged region is insensitive to the choice of lambda1 and lambda2, and the zero and non-zero elements under DSS-FGL tend to be quite stable as the shrinkage reaches a certain point.

So to start off we will set the parameters as follows, and will demonstrate the stability of the results in a later section.

lam1 <- 1

lam2 <- 1

v1 <- 1

lam.eff <- lam1 + c(1:10) * 5

v0s <- lam1/lam.eff

set.seed(1219)
G <- 4
p <- 51
penalty <- "fused"
lam1 <- 1
lam2 <- 1
v1 <- 1
lam.eff <- lam1 + c(1:10) * 5
v0s <- lam1/lam.eff

v0s
##  [1] 0.16666667 0.09090909 0.06250000 0.04761905 0.03846154 0.03225806
##  [7] 0.02777778 0.02439024 0.02173913 0.01960784

2.2.2 Run ssjgl function

The ssjgl() function takes input similar to the JGL package, with a few additional parameters as specified above.

fit1 = ssjgl(Y=dat_mat_list,penalty=penalty,lambda0=1, lambda1=lam1,lambda2=lam2, v1 = v1, v0s = v0s, a=1, b=p, doubly=TRUE, normalize=FALSE)

2.2.2.1 Diagnostic

First, we can plot the solution path and estimated precision matrices of SSJGL, as done in the paper. Essentially we are plotting the estimated partial correlations between all nodes over the iterations of the EM algorithm.

We can see that over the iterations many of the partial correlations shrink down to 0 while the non-zero elements stabilize, which helps us feel confident that the EM algorithm converged.

# Get plot titles 
mains <- c(paste0("Class 1"), 
           paste0("Class 2"), 
           paste0("Class 3"),
           "Class 4")

plot_path(lam1/v0s, fit1, thres = 0, normalize = T, xlab = expression(lambda[1]/v[0]), ylab = "Partial correlations",  main = mains, par = c(2, 2),
          position = "none")

2.2.2.2 Plotting estimated partial correlations

Now we can plot the estimated partil correlations from our DSS-FGL solution.

We see that for each class, there are 3 big “chunks” within the heatmap, which correspond with the 3 classes of metabolites. The patterns do look largely similar.

par(mfrow = c(2, 2))
for(i in 1:length(dat_mat_list)){
  corrplot::corrplot(-cov2cor(fit1$thetalist[[10]][[i]]), method="color", diag=F, is.corr=F)
}

#I am also going to specify colors based on the three metabolite groups and whether or not an edge indicates positive or negative correlation.
# specify colors for vertices (groups of metabolites in our case)

color_opts <- RColorBrewer::brewer.pal(3, "Dark2")
color_list <- data.frame(class = met_group) %>%
  mutate(color = case_when(
    class == "Amino Acids" ~ color_opts[1],
    class == "Acyl carnitines" ~ color_opts[2],
    class == "Other" ~ color_opts[3],
    TRUE ~ NA_character_
    )
  )

# color and labels for edges
edge_palette <- brewer.pal(n = 8, name = "RdBu")[c(8,1)]
edge_labels <- c("Positive", "Negative")

2.2.3 DSS-FGL Results

Finally we will extract estimated covariance matrices and prepare for visualizing the graphical model using the igraph package.

# extract all estimated covariance matrices from result
inv_covar_matrices <- fit1$thetalist[[length(v0s)]]
names(inv_covar_matrices) <- ancestry_groups

#now use `graph_from_adjacency_matrix()` function from igraph to create igraph graphs from adjacency matrices

#again since we have a list covariance matrices I use the map() function to apply igraph::graph_from_adjacency_matrix() over the list

graph_list <- map(ancestry_groups,
                  ~graph_from_adjacency_matrix(
                    -cov2cor(inv_covar_matrices[[.x]]),
                    weighted = T,
                    mode = "undirected",
                    diag = FALSE
                  ))
names(graph_list) <- ancestry_groups

2.2.3.1 DSS-FGL Class 1

plot_jgl(graph_list[[1]],  multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 1, DSS-FGL")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

2.2.3.2 DSS-FGL Class 2

plot_jgl(graph_list[[2]], multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 2, DSS-FGL")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

2.2.3.3 DSS-FGL Class 3

plot_jgl(graph_list[[3]], multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 3, DSS-FGL")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

2.2.3.4 DSS-FGL Class 4

plot_jgl(graph_list[[4]], multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 4, DSS-FGL")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

2.3 Running DS-GGL

Now we will run the doubly spike-and-slab group graphical lasso (DS-GGL). All that needs to be changed in the code is to specify the “group” penalty.

penalty = "group"

fit_group = ssjgl(Y=dat_mat_list,penalty=penalty,lambda0=1, lambda1=lam1,lambda2=lam2, v1 = v1, v0s = v0s,   a=1, b=p, doubly=TRUE, normalize=FALSE)

2.3.0.1 Diagnostic

As above, we can plot the solution paths to get a sense of if the solution converged.

# Get plot titles 
mains <- c(paste0("Class 1"), 
           paste0("Class 2"), 
           paste0("Class 3"),
           "Class 4")

plot_path(lam1/v0s, fit_group, thres = 0, normalize = T, xlab = expression(lambda[1]/v[0]), ylab = "Partial correlations",  main = mains, par = c(2,2),
          position = "none")

2.3.0.2 Plotting estimated partial correlations

Now we can plot the estimated partil correlations from our DSS-FGL solution.

We see that for each class, there are 3 big “chunks” within the heatmap, which correspond with the 3 classes of metabolites. The patterns do look largely similar.

par(mfrow = c(2, 2))
for(i in 1:length(dat_mat_list)){
  corrplot::corrplot(-cov2cor(fit_group$thetalist[[10]][[i]]), method="color", diag=F, is.corr=F)
}

2.3.1 DSS-FGL Results

Extract estimated covariance matrices and prepare for visualization.

# extract all estimated covariance matrices from result
inv_covar_matrices2 <- fit_group$thetalist[[length(v0s)]]
names(inv_covar_matrices2) <- ancestry_groups

graph_list_group <- map(ancestry_groups,
                  ~graph_from_adjacency_matrix(
                    -cov2cor(inv_covar_matrices2[[.x]]),
                    weighted = T,
                    mode = "undirected",
                    diag = FALSE
                  ))
names(graph_list_group) <- ancestry_groups

2.3.1.1 DSS-GGL Class 1

plot_jgl(graph_list_group[[1]],  multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 1, DSS-GGL")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

2.3.1.2 DSS-GGL Class 2

plot_jgl(graph_list_group[[2]], multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 2, DSS-GGL")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

2.3.1.3 DSS-GGL Class 3

plot_jgl(graph_list_group[[3]], multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 3, DSS-GGL")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

2.3.1.4 DSS-GGL Class 4

plot_jgl(graph_list_group[[4]], multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 4, DSS-GGL")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

3 Effect of parameters

To demonstrate the effect of various initial parameter choices I ran several additional models with “extreme” parameter choices.

3.1 a = b = 1

i = 1
test_graph <- graph_from_adjacency_matrix(-cov2cor(fit_group_smallp$thetalist[[length(v0s)]][[i]]),
                                    weighted = T,
                    mode = "undirected",
                    diag = FALSE)
plot_jgl(test_graph,  multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 1, DSS-GGL, a = b = 1")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

3.2 a = b = p

test_graph2 <- graph_from_adjacency_matrix(-cov2cor(fit_group_bigabigb$thetalist[[length(v0s)]][[i]]),
                                    weighted = T,
                    mode = "undirected",
                    diag = FALSE)
plot_jgl(test_graph2,  multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = "Class 1, DSS-GGL, a = b = p")
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

3.3 \(\lambda_0\) = 0

test_graph3 <- graph_from_adjacency_matrix(-cov2cor(fit_group_lam0_0$thetalist[[length(v0s)]][[i]]),
                                    weighted = T,
                    mode = "undirected",
                    diag = FALSE)
plot_jgl(test_graph3,  multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = bquote("Class 1, DSS-GGL"~lambda[0]==0))
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

3.4 \(\lambda_0\) = 0.01

test_graph3 <- graph_from_adjacency_matrix(-cov2cor(fit_group_lam0_01$thetalist[[length(v0s)]][[i]]),
                                    weighted = T,
                    mode = "undirected",
                    diag = FALSE)
plot_jgl(test_graph3,  multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = bquote("Class 1, DSS-GGL"~lambda[0]==0.1))
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

3.5 \(\lambda_0\) = 10

test_graph3 <- graph_from_adjacency_matrix(-cov2cor(fit_group_lam0_10$thetalist[[length(v0s)]][[i]]),
                                    weighted = T,
                    mode = "undirected",
                    diag = FALSE)
plot_jgl(test_graph3,  multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = bquote("Class 1, DSS-GGL"~lambda[0]==10))
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

3.6 \(\lambda_1\) = 0

No edges

test_graph4 <- graph_from_adjacency_matrix(-cov2cor(fit_group_lam1_0$thetalist[[length(v0s)]][[i]]),
                                    weighted = T,
                    mode = "undirected",
                    diag = FALSE)
# plot_jgl(test_graph4,  multiplier = 3, 
#          vertex.color = color_list$color,
#          vertex.label.color = color_list$color,
#          main = "Class 1, DSS-GGL, zero lambda_1")
# legend("bottomleft", inset=.02, title="Metabolite class",
#    unique(color_list$class), fill=color_opts, cex=0.8)
# legend("topleft", inset=.02, title="Edge",
#    edge_labels, fill=edge_palette, cex=0.8)

3.7 \(\lambda_1\) = 0.1

No edges

# test_graph4 <- graph_from_adjacency_matrix(-cov2cor(fit_group_lam1_01$thetalist[[length(v0s)]][[i]]),
#                                     weighted = T,
#                     mode = "undirected",
#                     diag = FALSE)
# plot_jgl(test_graph4,  multiplier = 3,
#          vertex.color = color_list$color,
#          vertex.label.color = color_list$color,
#          main = "Class 1, DSS-GGL, lambda_1 = 0.1")
# legend("bottomleft", inset=.02, title="Metabolite class",
#    unique(color_list$class), fill=color_opts, cex=0.8)
# legend("topleft", inset=.02, title="Edge",
#    edge_labels, fill=edge_palette, cex=0.8)

3.8 \(\lambda_1\) = 10

test_graph4 <- graph_from_adjacency_matrix(-cov2cor(fit_group_lam1_10$thetalist[[length(v0s)]][[i]]),
                                    weighted = T,
                    mode = "undirected",
                    diag = FALSE)
plot_jgl(test_graph4,  multiplier = 3, 
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = bquote("Class 1, DSS-GGL"~lambda[1]==10))
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

3.9 \(\lambda_2\) = 0

No edges

test_graph4 <- graph_from_adjacency_matrix(-cov2cor(fit_group_lam2_0$thetalist[[length(v0s)]][[i]]),
                                    weighted = T,
                    mode = "undirected",
                    diag = FALSE)
# plot_jgl(test_graph4,  multiplier = 3,
#          vertex.color = color_list$color,
#          vertex.label.color = color_list$color,
#          main = "Class 1, DSS-GGL, zero lambda_1")
# legend("bottomleft", inset=.02, title="Metabolite class",
#    unique(color_list$class), fill=color_opts, cex=0.8)
# legend("topleft", inset=.02, title="Edge",
#    edge_labels, fill=edge_palette, cex=0.8)

3.10 \(\lambda_2\) = 0.1

test_graph4 <- graph_from_adjacency_matrix(-cov2cor(fit_group_lam2_01$thetalist[[length(v0s)]][[i]]),
                                    weighted = T,
                    mode = "undirected",
                    diag = FALSE)
plot_jgl(test_graph4,  multiplier = 3,
         vertex.color = color_list$color,
         vertex.label.color = color_list$color,
         main = bquote("Class 1, DSS-GGL"~lambda[2]==0.1))
legend("bottomleft", inset=.02, title="Metabolite class",
   unique(color_list$class), fill=color_opts, cex=0.8)
legend("topleft", inset=.02, title="Edge",
   edge_labels, fill=edge_palette, cex=0.8)

4 Coments: comparing to frequentist method JGL

Check out my tutorial on the frequentist method “Joint Graphical Lasso” (JGL) using the same metabolite data in this tutorial to compare the analysis results.

A few observations I have:

  • While SSJGL can take some time to run depending on your data size and parameters, I found it to be much faster than having to do cross-validation to tune the penalty terms using JGL. SSJGL does not require cross-validation or tuning, because the results are quite stable to the choice of input parameters.

  • I also found the final result of SSJGL to be much more sparse compared to JGL. For example in this metabolite dataset, using DSS-GGL to jointly estimate the graph structure, Class 1 had 138 edges. This contrasts to the final tuned GGL graph, which had 240 edges.

    • According to the SSJGL paper, graph selection using AIC tends to favor larger models, which is exactly what we see happening here.

  • I also found the results of SSJGL to be stable no matter what starting parameters I threw at it, with the number of the edges in the solution ranging from 102-148 edges. This contrasts to when I threw all kinds of lambdas into JGL and the solution for this same dataset ranged from 44 edges (with large penalty parameter lambdas) to 692 edges (with small lambdas).